TL;DR
This paper introduces a novel 3D-Mask RCNN framework for detecting and segmenting breast masses in Digital Breast Tomosynthesis (DBT) images. The 3D approach significantly outperforms its 2D counterparts, particularly in challenging cases like dense breasts and irregular mass shapes, paving the way for more accurate and efficient breast cancer screening.
What Problem Does This Paper Address?
Digital Breast Tomosynthesis (DBT) is a promising advancement in breast cancer screening, offering quasi-3D imaging that helps overcome the limitations of traditional mammography, especially in dense breast tissue where overlapping structures can obscure suspicious lesions. However, the increased volume of data from DBT also leads to a higher workload for radiologists. To address this, computer-aided diagnosis (CAD) systems are crucial. While existing CAD systems have explored 2D analysis of DBT slices, the potential of leveraging the full 3D information for improved mass detection and segmentation has remained largely unexplored. This paper aims to bridge this gap by developing and evaluating a 3D deep learning approach for mass detection and segmentation in DBT.
Methods
The core of this research is the development of a 3D-Mask Region-Based Convolutional Neural Network (3D-Mask RCNN). This framework builds upon the successful Mask RCNN architecture, which is known for its ability to perform both object detection and instance segmentation. The key innovation here is the adaptation of Mask RCNN to handle 3D volumetric data from DBT.
To manage the computational demands of 3D convolutions, the researchers processed the DBT images in smaller 3D patches (256x256x64 voxels). The network utilizes a Residual Network (ResNet)-Feature Pyramid Network (FPN) backbone to extract features at various scales. A Region Proposal Network (RPN) generates candidate bounding boxes, and RoIAlign is employed for precise feature map alignment. The detection branch performs bounding box regression and classification, while the mask branch generates pixel-level segmentation masks.
For comparison, the authors also implemented and evaluated 2D-Mask RCNN and Faster RCNN models, which operate on individual 2D slices of the DBT data.
Datasets
The study utilized a dataset of 364 DBT samples collected from Fudan University Affiliated Cancer Center. This dataset was divided into a training set (n=201) and a testing set (n=163). The paper provides a detailed breakdown of the patient characteristics, including age, histological type, mass type, breast density, and tumor maximum diameter, in Table 1. No new benchmarks were introduced; the focus was on evaluating existing architectures on DBT data.
Key Insights and Ideas
- 3D is better than 2D for DBT: The most significant insight is that leveraging the full 3D volumetric information of DBT, as done with the 3D-Mask RCNN, leads to demonstrably superior performance in mass detection and segmentation compared to methods that analyze 2D slices.
- Dense breasts are a key challenge, and 3D helps: The 3D approach showed a marked improvement in detecting masses in dense breasts, a known challenge for 2D mammography and potentially for 2D DBT analysis. This suggests that the volumetric context is crucial for disambiguating lesions in dense tissue.
- Irregular shapes and specific age groups are harder: The study identified that certain patient subgroups, such as those with irregular mass shapes and individuals in the 40-49 age range, presented greater detection challenges. The 3D-Mask RCNN showed significant improvements in these specific, more difficult cases.
- Segmentation benefits from 3D too: Beyond detection, the 3D-Mask RCNN also achieved superior performance in segmenting the precise boundaries of breast masses, indicating its utility for accurate characterization.
Key Results
- Improved Detection Accuracy: At a 90% sensitivity, the 3D-Mask RCNN achieved significantly fewer false positives (FPs) per lesion (0.8 FPs/lesion) and per breast (0.83 FPs/breast) compared to 2D-Mask RCNN (1.3 FPs/lesion, 1.24 FPs/breast) and Faster RCNN (2.37 FPs/lesion, 2.38 FPs/breast).
- Subgroup Performance: The 3D-Mask RCNN demonstrated statistically significant improvements (p < 0.05) over 2D methods for patients aged 40-49, those with malignant tumors, spiculate and irregular masses, and dense breasts.
- Superior Segmentation: For lesion segmentation, the 3D-Mask RCNN achieved an average precision (AP) of 0.934 and a false negative rate (FNR) of 0.053, outperforming the 2D methods.
- Qualitative Improvements: Visual examples (Figures 7-9) clearly illustrate the 3D-Mask RCNN's ability to better delineate masses, especially in challenging dense breast cases, and its reduced tendency for false positives compared to 2D methods.
Limitations and Future Work
The authors acknowledge several limitations:
- Sample Size: The study's sample size, while substantial, is considered relatively small, particularly for subgroup analyses. Future work should involve larger cohorts to further refine the findings.
- No Transfer Learning: The study did not incorporate transfer learning from mammography datasets. The authors suggest that future research could explore this to potentially enhance accuracy and segmentation performance.
- Patch-Based Processing: To manage memory constraints, the models were trained and tested on image patches. Future studies could investigate methods for processing entire DBT images to avoid potential patch-boundary artifacts.
The authors propose that future work should focus on these areas to further advance the capabilities of 3D deep learning for breast cancer detection and segmentation.
Summary updated